Separate large vocabularies (use custom validation) from small vocabularies and subsets (use codegen)

[dbirman]

Copying summary info from discussion w/ David + Saskia

1. check in large vocabs to aind-data-schema-models in CSV (or maybe parquet, they are big)
2. make subsets by creating separate subset files (e.g. lens_manufacturers.csv)
3. don't create all the pydantic classes, instead make sets that should hopefully load faster
4. make a library of custom validators in aind-data-schema-models
5. make use of those custom validators in aind-data-schema and decorate fields that use them with json_schema_extra
6. in metadata-entry, pull vocabs from docdb intelligently by looking at json_schema_extra

[dbirman]

Additional discussion w/ Jon and Bruno:

• For large voabs we can potentially validate at test time against the external registry
• Small vocabs should keep them internal and copied to docdb